remote sensing
Article
Estimating Rice Agronomic Traits Using Drone-Collected Multispectral Imagery
Dimitris Stavrakoudis1,2,* , Dimitrios Katsantonis1, Kalliopi Kadoglidou1, Argyris Kalaitzidis1and Ioannis Z. Gitas2
1 Hellenic Agricultural Organization—“DEMETER”, Institute of Plant Breeding and Genetic Resources, Thermi-Thessalonikis, Ellinikis Georgikis Scholis, GR-57001, Greece; dikatsa@cerealinstitute.gr (D.K.);
kadoglidou@ipgrb.gr (K.K.); kalaitzidis@cerealinstitute.gr (A.K.)
2 Laboratory of Forest Management and Remote Sensing, School of Forestry and Natural Environment, Aristotle University of Thessaloniki, P.O. Box 248, GR-54124, Greece; igitas@for.auth.gr
* Correspondence: jstavrak@auth.gr; Tel.: +30-2310-992689
Received: 12 February 2019; Accepted: 26 February 2019; Published: 6 March 2019 Abstract:The knowledge of rice nitrogen (N) requirements and uptake capacity are fundamental for the development of improved N management. This paper presents empirical models for predicting agronomic traits that are relevant to yield and N requirements of rice (Oryza sativaL.) through remotely sensed data. Multiple linear regression models were constructed at key growth stages (at tillering and at booting), using as input reflectance values and vegetation indices obtained from a compact multispectral sensor (green, red, red-edge, and near-infrared channels) onboard an unmanned aerial vehicle (UAV). The models were constructed using field data and images from two consecutive years in a number of experimental rice plots in Greece (Thessaloniki Regional Unit), by applying four different N treatments (C0: 0 N kg·ha−1, C1: 80 N kg·ha−1, C2: 160 N kg·ha−1, and C4:
320 N kg·ha−1). Models for estimating the current crop status (e.g., N uptake at the time of image acquisition) and predicting the future one (e.g., N uptake of grains at maturity) were developed and evaluated. At the tillering stage, high accuracies (R2≥0.8) were achieved for N uptake and biomass.
At the booting stage, similarly high accuracies were achieved for yield, N concentration, N uptake, biomass, and plant height, using inputs from either two or three images. The results of the present study can be useful for providing N recommendations for the two top-dressing fertilizations in rice cultivation, through a cost-efficient workflow.
Keywords: rice agronomic traits; multispectral UAV imagery; nitrogen uptake; nitrogen concentration; yield; aboveground biomass; multiple linear regression modeling; lasso input selection
1. Introduction
Rice (Oryza sativaL.) is the second most cultivated cereal crop and the most consumed staple food in the world, since more than three billion people rely on rice as their primary source of food. Although it is predominant in Asia, rice has also been cultivated in Europe since the 15th century, mainly in Mediterranean countries including Italy, Spain, Portugal, Greece, and France (FAO database 2018).
Rice is cultivated under a wide range of ecosystems, but more than 90% of the world’s rice production is harvested from irrigated or rainfed lowland rice fields [1]. Thus, increase in rice production is needed if the increased demand from population growth is to be met.
The use of nitrogen (N) fertilizer has changed the global N cycle markedly and has been causing various negative environmental consequences, such as eutrophication of surface water, global warming, and ozone layer depletion [2–4]. Modern production agriculture requires efficient, sustainable, and environmentally-sound management practices. N is a key factor in achieving optimum lowland
Remote Sens.2019,11, 545; doi:10.3390/rs11050545 www.mdpi.com/journal/remotesensing
rice grain yields [5]. It is the nutrient input normally required in large quantities for achieving high yields, but soils under these conditions are saturated, flooded, and anaerobic and, therefore, N use efficiency is low [6]. However, more than 50% of the applied N is not assimilated by the rice plant and it is lost through different mechanisms including ammonia volatilization, surface runoff, nitrification-denitrification, and leaching [7–9]. N2O is the main greenhouse gas (GHG) related to agricultural soil emissions, essentially due to microbial transformation of nitrogen in the soil. This concerns N mineral fertilizers, manure spreading, and N from crop residues incorporated into the soil or lixiviation of surplus nitrogen. N2O has high global warming potential (298 times higher than CO2) and it should be minimized to reduce agricultural GHG emissions in total. The application of mineral N in the form of chemical fertilizers can also increase the N2O emissions [10]. A recent study estimated the seasonal direct emission of N2O from the paddy rice system in China (a country accounting for approximately 30% of the global rice production) to be 31.1 Gg N2O-N for 2014, analyzing data from multiple studies [11]. Another recent study reported that intermittent flooding practices in rice can significantly increase N2O emissions [12], although there is an active debate on this issue [13,14].
Therefore, N fertilizer management strategies that increase crop productivity and N use efficiency, while reducing negative environmental consequences, have to focus on parameters such as optimum time, rate, and spatial distribution methods that synchronize plant N requirements with N supply, in order to reduce N losses and maximize uptake of applied N in the crop [15]. Legislative measures adopted to comply with the Directive 91/676/EEC concerning the “Protection of Waters against Pollution caused by Nitrates from Agricultural Sources,” in some cases did not obtain the expected results and are not always accepted (or complied with) by farmers [16,17].
A key contributor to yield increases will be the efficient and effective use of nitrogen (N) fertilizer, which is relatively low in irrigated rice, because of rapid N losses from volatilization and denitrification in the soil-floodwater system [18]. The “4R” nutrient stewardship—applying the right nutrient source at the right rate, at the right time, and in the right place—is an innovative approach in fertilizer management [19]. Precision agriculture also has a positive impact on farm productivity and economics, as it provides higher or equal yields with lower production cost than conventional practices. The adoption of precision agriculture techniques for N management has the potential for improving agronomic, economic, and environmental efficiency in the use of such input [20].
Regarding the rest of the fertilization practices generally followed in Europe, phosphorous and potassium are supplied in the pre-planting stage at 50–70 kg·ha−1and 100–150 kg·ha−1, respectively.
The first fertilization intervention usually provides a nitrogen–phosphorous–potassium complex; it is carried out before the field flooding. The second supply is usually applied when rice plants are at the 3–5-leaf stage, at the beginning of tillering. A third fertilization can sometimes occur during panicle initiation. Sulfur can be applied using sulfuric fertilizers (NH4or K) and zinc is generally needed in soils with high pH, whereas calcium is needed in pathogenic soils with high salinity. The latter elements are used in Europe occasionally or after deficiencies [21].
A major challenge in N management through precision agriculture techniques is the accurate prediction and mapping of plant agronomic traits related to plant nutrient status and—most importantly—to N content. The traditional soil-based testing methods widely used for upland crops are not suitable for N recommendations in rice fields due to the flooding and complexity of N cycling in the rice paddy’s soil during the growing period [22]. Even before flooding, the available soil-N test is not accurate enough to lead to confident recommendations for N fertilizer applications [23].
Remote sensing technologies offer a viable alternative for deriving precise N recommendations in rice fields through the dynamic non-destructive estimation of plant N status throughout the growing season, along with predictions of other agronomic traits of interest (e.g., yield). Coupled with the rapidly advancing technology of unmanned aerial vehicle (UAV) platforms, they are nowadays starting to offer cost-effective solutions to various aspects of crop monitoring and sustainable crop management [24–27].
Remote Sens.2019,11, 545 3 of 25
A few studies have proposed employing the traditional approach of active canopy sensors for estimating yield-related agronomic traits of rice plants, which can help in supporting precision N management [28–35]. For active field applications, these sensors could be installed on the machinery for real-time sensing, but such systems are not common in rice cultivation [36]. Most importantly, their use does become cost-inefficient (time and energy consuming) if the tractor must enter the rice paddy just for monitoring the N status, at growth stages other than the appropriate for fertilizer application or for weed and pest management. A substantial number of studies have also investigated the relationship between narrow-band vegetation indices (VIs) and various agronomic traits of rice plants [22,37–54].
The VIs are calculated from the spectral signatures obtained from handheld spectrometers, which are portable non-imaging hyperspectral sensors acquiring single spectral signatures from a small (typically circular) surface over the canopy. Most of these studies construct empirical models through linear regression for each possible pair of agronomic traits versus VIs, thus identifying the individual VIs that are highly correlated with each agronomic trait, but using the whole hyperspectral signatures as input to multivariable N status prediction models has also been proposed [45,48].
The results of the aforementioned studies are important, since they have identified the spectral wavelengths and/or VIs that are highly correlated with yield-related traits (e.g., [32,33,46,52]) or even proposed N recommendation methodologies based on remotely sensed data [28]. However, canopy-level spectral data from spectrometers becomes inappropriate for fine-scale precision farming, since that would require an extremely time-consuming and cost-inefficient dense sampling. In addition, the calculation of optimized VIs requires the sensor to incorporate bands at very specific wavelengths, which are typically not available in most multispectral sensors. It is true that compact hyperspectral sensors that can be mounted onboard UAVs (or movable structures above the canopy) have been developed the last few years and have been successfully employed for estimating rice agronomic traits [55–57]. Yet, the use of these systems is still limited in precision agriculture applications, their cost is relatively high, and processing such large volumes of data is quite challenging, although they have great potential for being operationally employed in the future.
Multispectral imaging sensors probably constitute the most appropriate system for precision agriculture applications in rice cultivation. Traditionally, satellite imagery has been employed for monitoring rice growth and agronomic traits. Medium-resolution satellite imagery (spatial resolution of 20–30 m) has been employed for providing rather coarse estimations of rice crop traits [58–61] or has been incorporated within rice growth simulation models [62]. High-resolution synthetic aperture radar (SAR) data have been employed for estimating morphological traits (most notably height) [63].
Recently, Sentinel-2 data have also been employed for estimating leaf area index (LAI) and plant N concentration of rice fields through empirical regression modeling, which were then used for producing N nutritional index maps [64]. Finally, a few studies have used commercial high-resolution (spatial resolution less than 10 m) satellite imagery to estimate N status [36,65,66] or within-field variability [61].
However, the most important drawback of high-resolution satellite images is their high cost for real-life applications, as almost all vendors enforce a large minimum area that can be ordered for a single image.
If multiple images within the season are required for providing N recommendations and taking into consideration that rice paddies in a single cultivation have variable sowing dates—typically up to 30 days difference (due to genotypic differences in plant stages etc.)—the cost increases exponentially.
With the rapid development of UAV platforms (load capacity and flying autonomy), compact and lightweight multispectral sensors onboard UAVs seems to be the most direct and cost-efficient approach to precision agriculture in rice today. However, very few studies have been published that exploit such systems for estimating agronomic traits of rice plants. Initially, plain RGB cameras were used for estimating N status of rice plants [67,68], but these cameras are difficult to calibrate for different lighting conditions and they do not incorporate a near-infrared (NIR) channel that is required for the calculation of most VIs. Modifying the camera with a polarizing filter could solve the latter problem [69], but the sensor’s spectral response is still not appropriate for accurate calculation of most VIs, which require specific ranges of wavelengths. Good correlations of specific VIs calculated from
multispectral sensors (RGB plus NIR channel) and measurements from chlorophyll meters (SPAD meter) have been reported [70,71]. However, estimating leaf N content from SPAD readings is not straightforward, since significant variability has been observed between SPAD values and both leaf chlorophyll [72–74] and leaf N [75–77] content. Finally, Lu et al. [78] has tested the potential of a 5-band multispectral sensor (RGB plus red-edge (RE) and NIR channels) onboard a UAV for estimating LAI and N content of rice plants, through simple linear regression. Medium to high correlations were reported at the panicle growing stage—close to the second top-dressing fertilization (TdF)—with lower coefficient of determination (R2) values reported for the stem elongation stage (close to the first TdF).
However, their experimentation considered field data from a single growing season, and the lower R2 values reported for the stem elongation stage suggest that multiple remote sensing variables must be considered simultaneously for increasing the modeling accuracy.
The objective of the current paper is to present empirical models for predicting agronomic traits of rice plants through remotely-sensed multispectral imagery. Models for predicting plant height, aboveground biomass, N concentration, N uptake, grain yield, and harvest index have been constructed following a multiple linear regression approach at different growth stages (at tillering and at booting), using as input reflectance values and VIs obtained from a compact four-band sensor (green, red, narrow-band RE, and NIR channels) onboard a UAV. The models were constructed using field data and images from two consecutive years in a number of experimental rice plots in Greece, employing different N treatments. Models for both estimating the current crop status (e.g., N uptake at the time of image acquisition) and predicting the future one (e.g., N uptake of grains at maturity) were developed and evaluated. To the best of our knowledge, this is the first study that presents such a comprehensive analysis for estimating various yield-related agronomic traits of rice plants during the growing season from UAV-collected multispectral data.
2. Materials and Methods
2.1. Field Experiments
Two field experiments were conducted in two consecutive years, 2016 and 2017, at the Experimental Station of Kalochori, Thessaloniki, Greece (40◦36058.75”N, 22◦49051.16”E in WGS 84 spatial reference system). The size of each experimental plot area was 11 m2(5×2.5 m) and arranged in a randomized complete block design with five replications for each treatment (Figure1). The plots were fertilized with a novel bio-fertilizer (RBBf), developed at the Hellenic Agricultural Organization – DEMETER, in the framework of the H2020 project AGROCYCLE. The RBBf comprised 74% rice industrial co-products, chicken manure, zeolite, a compost accelerator by addingAspergilusspp.
(fungi), Bacillusspp. (bacteria), and larvae of the insectsHermetia illucens andCetonia aurata. All raw materials were placed into a custom-made automatic compost bin of 1.5 tons capacity and left for at least 40 days to complete the composting process (i.e., when the C/N ratio reached a value lower than 20). The experiment consisted of four fertilization treatments: 1) 0 N kg·ha−1—untreated control (C0); 2) 80 N kg·ha−1of RBBf (C1); 3) 160 N kg·ha−1of RBBf—standard local practice for rice cultivation (C2); and 4) 320 N kg·ha−1RBBf (C3). Seeds of the japonica type Greek commercial variety DION—belonging to the European Core Collection (http://tropgenedb.cirad.fr)—were directly seeded into the flooded plots on 2 June 2016 and 9 June 2017, respectively. The fertilization scheme followed the local and international standard practices for rice cultivation, where the amount is divided in three increments: 40% of the N–P–K are incorporated as basal before flooding, 40% is applied at the tillering stage, and similarly the rest (20%) at the panicle initiation stage. Besides fertilization, all the rest of the common practices were applied. The whole experiment was harvested when the plants reached the physiological maturity stage on October 6, 2016 and October 10, 2017, respectively.
Remote Sens.2019,11, 545 5 of 25
Remote Sens. 2018, 10, x FOR PEER REVIEW 4 of 26
leaf chlorophyll [72–74] and leaf N [75–77] content. Finally, Lu et al. [78] has tested the potential of a 5-band multispectral sensor (RGB plus red-edge (RE) and NIR channels) onboard a UAV for estimating LAI and N content of rice plants, through simple linear regression. Medium to high correlations were reported at the panicle growing stage—close to the second top-dressing fertilization (TdF)—with lower coefficient of determination (R2) values reported for the stem elongation stage (close to the first TdF). However, their experimentation considered field data from a single growing season, and the lower R2 values reported for the stem elongation stage suggest that multiple remote sensing variables must be considered simultaneously for increasing the modeling accuracy.
The objective of the current paper is to present empirical models for predicting agronomic traits of rice plants through remotely-sensed multispectral imagery. Models for predicting plant height, aboveground biomass, N concentration, Ν uptake, grain yield, and harvest index have been constructed following a multiple linear regression approach at different growth stages (at tillering and at booting), using as input reflectance values and VIs obtained from a compact four-band sensor (green, red, narrow-band RE, and NIR channels) onboard a UAV. The models were constructed using field data and images from two consecutive years in a number of experimental rice plots in Greece, employing different N treatments. Models for both estimating the current crop status (e.g., N uptake at the time of image acquisition) and predicting the future one (e.g., N uptake of grains at maturity) were developed and evaluated. To the best of our knowledge, this is the first study that presents such a comprehensive analysis for estimating various yield-related agronomic traits of rice plants during the growing season from UAV-collected multispectral data.
Figure 1. Location of the experimentation station (red rectangles) in Greece (top left), overview of whole Experimental Station in Kalochori (top right), and the setup of the experimental plots (bottom).
In the latter, the color of each plot denotes each treatment’s amount of nitrogen (N) applied (C0: 0 kg∙ha−1, C1: 80 kg∙ha−1, C2: 160 kg∙ha−1, and C3: 320 kg∙ha−1).
2. Materials and Methods
2.1. Field Experiments
Two field experiments were conducted in two consecutive years, 2016 and 2017, at the Experimental Station of Kalochori, Thessaloniki, Greece (40°36'58.75"N, 22°49'51.16"E in WGS 84
Figure 1. Location of the experimentation station (red rectangles) in Greece (top left), overview of whole Experimental Station in Kalochori (top right), and the setup of the experimental plots (bottom).
In the latter, the color of each plot denotes each treatment’s amount of nitrogen (N) applied (C0:
0 kg·ha−1, C1: 80 kg·ha−1, C2: 160 kg·ha−1, and C3: 320 kg·ha−1).
Soil sampling was carried out prior to flooding, with three samples being collected as bulk from different field points according to standard soil field sampling methods. The soil analysis average results were sand = 30%, silk = 18%, loam = 52%, pH = 7.6, organic matter = 2%, N = 0.05%, P = 0.012%, K = 0.24%, Ca = 0.2%, and Zn = 0.012%. However, no significant variations within the different samples collected from the whole experimentation area were observed. Standard cultural practices for rice cultivation were conducted including harrowing, tillage, and lazier leveling. The plot embankments were made by hand, as were the RBBf application and sowing. Weed control was performed chemically in the whole period of herbicide application according to the local standard practices. The irrigation water is of a very good quality (0.7 dS) in the area, without any traces of agrochemical residuals since it is checked very frequently from the local irrigation organization. After harvesting, the whole plants were removed from the paddies, whereas the crop residuals—including 2–3 cm culm and the roots—remained in the soil and were incorporated for decomposition until the next year. In the second year of the experimentation, we chose not to alter the position of each paddy-treatment, to avoid mixing of the different treatments throughout the experimentation.
During the experimentation period, the following traits were assessed at the appropriate BBCH-scale stages [79,80]: (1) plant height (PH; cm), by measuring the main stem of 30 random rice plants per plot; (2) total biomass (BT; tn·ha−1), by cutting 0.25m2 square plots of the whole aboveground plant biomass and weighing it after 48 h of air oven drying at 70◦C until constant weight; (3) N concentration (NC; %), determined by the macro-Kjeldahl procedure [81]; (4) N uptake (NU; kg·ha−1), estimated as total biomass ×N concentration [36]; (5) grain yield (Yield; tn·ha−1);
and (6) harvest index (HI), estimated as grain yield / (total dry matter + grain yield). Harvest index is the ratio of grain yield to total biomass and is considered as a measure of biological success in partitioning assimilated photosynthate to the harvestable product [82]. The traits that change during the season (PH, BT, NC, and NU) were measured at BBCH scale: stage 25 (before the first TdF), stage
45 (at booting, approximately 5 days before the second TdF), and stage 99 (harvesting). The timings of the first two measurements coincided with the first and last UAV image acquisitions (see next subsection). At harvesting, NC and NU of stem and leaves only (NCSL and NUSL, respectively) and of grain (NCG and NUG, respectively) were also assessed. Similarly, biomass of stem and leaves only (BSL) was considered as an additional independent trait. Conversely, PH at maturity was not considered, since the parameter does not change after the full heading stage (BBCH 59 to 60). Table1 summarizes all agronomic traits considered in this study, which are used as dependent variables of the predictive models.
Table 1.Summary description of the rice agronomic traits considered in this study.
Abbreviation Description Units
PH25 Plant height at BBCH 25 cm
PH45 Plant height at BBCH 45 cm
BT25 Total aboveground biomass at BBCH 25 tn·ha−1 BT45 Total aboveground biomass at BBCH 45 tn·ha−1 BT99 Total aboveground biomass at maturity tn·ha−1 BSL99 Biomass of stem and leaves at maturity tn·ha−1
Yield Yield tn·ha−1
HI Harvest index —
NC25 Plant N concentration at BBCH 25 %
NC45 Plant N concentration at BBCH 45 %
NC99 Plant N concentration at maturity %
NCSL99 N concentration of stem and leaves at
maturity %
NCG N concentration of grains at maturity %
NU25 Plant N uptake at BBCH 25 kg·ha−1
NU45 Plant N uptake at BBCH 45 kg·ha−1
NU99 Plant N uptake at maturity kg·ha−1
NUSL99 N uptake of stem and leaves at maturity kg·ha−1
NUG N uptake of grains at maturity kg·ha−1
2.2. UAV Imagery and Preprocessing
We used the Parrot®Sequoia™ multispectral imaging sensor (Parrot Drones S.A.S, Paris, France), which captures the reflected light at four spectral bands with a field of view of 70.6◦: green (G; 550 nm;
40 nm bandwidth), red (R; 660 nm; 40 nm bandwidth), red-edge (RE; 735 nm; 10 nm bandwidth), and near-infrared (NIR; 790 nm; 40 nm bandwidth). The camera was mounted on a DJI Phantom 4 quadcopter (DJI, Shenzhen, China) and it was also equipped with an irradiance sensor for measuring incident light, in order to correct for variable lighting conditions during the flight. Prior to each flight, ground images from a calibration target (AIRINOV Aircalib; AIRINOV, Paris, France) were acquired, in order to derive accurate reflectance values. The latter was a polyvinyl chloride (PVC) board with a gray target area silkscreen printing and ArUco tags for the albedo measurements of each band, which have been measured specifically for the Sequoia sensor. The flight plan was created automatically through the Atlas Flight application (MicaSense, Inc., Seattle, USA), choosing 80% overlap for both front-lap and side-lap, in order to assure an accurate orthomosaic creation. The single images were subsequently combined to create an orthophotomosaic using Pix4Dmapper Pro (Pix4D S.A., Lausanne, Switzerland), which employs an advanced structure from motion (SfM) workflow to derive accurate orthophotomosaics in absolute reflectance values. The derivation of absolute reflectance values is achieved by Pix4D taking into consideration both the calibration target images and the readings from the irradiance sensor in an automated workflow, without the need for the user to calibrate or otherwise process the original images. An overview of the typical acquisition procedures and required processing of UAV imagery can be found in [83].
Remote Sens.2019,11, 545 7 of 25
During each growing season, three images were acquired over the study area (Table2). The drone was flown at a height of 30 m above ground level, resulting in an orthophotomosaic with spatial resolution of approximately 2.5 cm. A total of 35 VIs were calculated from each image (Table3), which have been considered in rice-related studies for the estimation of agronomic traits. Subsequently, the area within each experimental plot was manually delineated (approximately 10 m2), taking care not to include the plots’ boundaries (the paddy’s embankments). For each plot, the average value of all pixels within the delineated area was calculated for each camera band and VI, which were subsequently used for the analysis. The final dataset comprised 40 samples, since the experiment comprised 20 plots (five replications for each of the four N treatments) and measurements were acquired for the two years of experimentation (2016 and 2017).
Table 2.Dates, equivalent days after sowing (DAS), and BBCH stage code for the major treatments, field data collection, and unmanned aerial vehicle (UAV) image acquisitions performed in the experimental plots during the two years.
Treatment or Data Acquisition Date DAS BBCH
2016
Basal fertilization 2 June 2016 0 —
Sowing 2 June 2016 0 —
1st image acquisition and field data collection 5 July 2016 33 25
1st TdF 8 July 2016 36 26
2nd image acquisition and field data collection 15 July 2016 43 31 3rd image acquisition and field data collection 17 August 2016 76 45
2nd TdF 22 August 2016 81 49
Harvesting 6 October 2016 126 99
2017
Basal fertilization 8 June 2017 −1 —
Sowing 9 June 2017 0 —
1st image acquisition and field data collection 16 July 2017 37 25
1st TdF 18 July 2017 39 26
2nd image acquisition and field data collection 25 July 2017 46 31 3rd image acquisition and field data collection 18 August 2017 70 45
2nd TdF 23 August 2017 75 49
Harvesting 10 October 2017 123 99
Table 3.Summary of vegetation indices considered in this study. Camera channels are reported as G:
green, R: red, RE: red-edge, and NIR: near-infrared.
Acronym Name Formula Introduced in
DVI Difference Vegetation Index (VI) NIR−R [84]
NDVI Normalized difference VI (NIR−R)/(NIR+R) [85]
RVI Ratio VI (also simple ratio (SR)) NIR/R [86]
mSR Modified simple ratio (NIR/R−1)/p
(NIR/R+1) [87]
TNDVI Transformed NDVI p
(NIR−R)/(NIR+R) +0.5 [84]
RDVI Renormalized DVI (NIR−R)/p
(NIR+R) [88]
SAVI Soil-adjusted VI 1.5·(NIR−R)/(NIR+R+0.5) [89]
OSAVI Optimized SAVI 1.16·(NIR−R)/(NIR+R+0.16) [90]
MSAVI2 Modified SAVI 2 0.5·h2NIR+1−(2NIR+1)2−8√ NIR−Ri
[91]
gDVI Green DVI NIR−G [84]
gNDVI Green NDVI (NIR−G)/(NIR+G) [92]
gRDVI Green RDVI (NIR−G)/p
(NIR+G) [32]
mSRG Modified green simple ratio (NIR/G−1)/p
(NIR/G+1) [32]
GSAVI Green SAVI 1.5·(NIR−G)/(NIR+G+0.5) [93]
MGSAVI Modified GSAVI 0.5·
2NIR+1− q
(2NIR+1)2−8(NIR−G)
[32]
NGI Normalized green index G/(NIR+RE+G) [93]
GWDRVI Green wide dynamic range VI (0.12·NIR−G)/(0.12·NIR+G) [32]
Table 3.Cont.
Acronym Name Formula Introduced in
CIG Green chlorophyll index NIR/G−1 [94]
CIRE Red-edge chlorophyll index NIR/RE−1 [94]
TCARI Transformed chlorophyll
absorption ratio index 3·[(RE−R)−0.2(RE−G)(RE/R)] [95]
MCARI1 Modified chlorophyll absorption
in reflectance index 1 1.2·[2.5(NIR−R)−1.3(NIR−G)] [96]
MCARI2 Modified chlorophyll absorption in reflectance index 2
1.5·[2.5(NIR−R)−1.3(NIR−G)]/ r
(2N IR+1)2−h6N IR−5√ Ri
−0.5 [96]
TCARI/ OSAVI TCARI to OSAVI TCARI/OSAVI [95]
REDVI Red-edge DVI NIR−RE [32]
NDRE Normalized difference red-edge
index (NIR−RE)/(NIR+RE) [97]
RERDVI Red-edge RDVI (NIR−RE)/p
(NIR+RE) [32]
NNIR Normalized NIR index NIR/(NIR+RE+G) [93]
REGNDVI Red-edge GNDVI (RE−G)/(RE+G) [32]
REWDRVI Red-edge wide dynamic range
VI (0.12·NIR−RE)/(0.12·NIR+RE) [32]
MSRRE Modified red-edge simple ratio (NIR/RE−1)/p
(NIR/RE+1) [32]
MEVI Modified enhanced VI 2.5·(NIR−RE)/(NIR+6RE−7.5G+1) [32]
RESAVI Red-edge SAVI 1.5·(NIR−RE)/(NIR+RE+0.5) [32]
MRESAVI Modified RESAVI
0.5·
2NIR+1−q(2NIR+1)2−8(NIR−RE)
[32]
MTCARI Modified TCARI 3·[(NIR−RE)−0.2(NIR−G)(NIR/RE)] [32]
MRETVI Modified RETVI 1.2·[1.2(NIR−G)−2.5(RE−G)] [32]
2.3. Regression Models
The union of the camera’s four bands and the 35 VIs were used as inputs for building the empirical models. The samples used for the analysis were the average value of the VIs and the camera’s bands—as mentioned in the previous section—together with the corresponding field data values (one of the traits for each model) for each plot and for the two years of experimentation. In order to maintain the models’ interpretability, we chose to create them through the simple least-square multiple linear regression [98] approach. All VIs were considered as potential inputs for the models, since N treatments affect, simultaneously, biomass, leaf area index, and chlorophyll content of all plants and, as such, any VI could possibly increase the accuracy of a model built for each trait. In other words, we cannota prioriassume that some VI is irrelevant with an agronomic trait. However, since the number of possible inputs was high (39 for the models built for the tillering stage and 78 or 117 for those built for the booting stage, as will be explained in the following), a subset of predictors had to be selected, in order to maintain the models’ interpretability and increase their generalization capabilities. Input selection was performed through the lasso (least absolute shrinkage and selection operator) [99], which is a regularization technique for performing linear regression. Lasso solves the minimization problem:
min
β0,β
1 2N
∑
N i=1yi−β0+xTi ·β 2
+λ
∑
P j=1βj
!
, (1)
whereNis the number of data samples,Pis the number of predictors (input variables),xiis the input vector (P-dimensional column vector) for theith data sample,yiis the corresponding output value,β is theP-dimensional column vector of linear coefficients,β0is the intercept term, andλis the positive regularization parameter. Lasso incorporates theL1norm ofβinto the minimization task, which forces a number of coefficients to become zero, thus identifying the redundant predictors. Prior to the analysis, the input data were standardized to zero mean and variance of one. The regularization parameterλ was determined through a five-fold cross-validation procedure, considering a geometric sequence of 100 values, with the largest value being the one that produces a null model. The largestλvalue such that the mean squared error (MSE) was within one standard error of the minimum MSE was ultimately selected. The predictors with non-zero coefficients were maintained and a multiple linear regression
Remote Sens.2019,11, 545 9 of 25
model was constructed considering those predictors only. To increase the model’s interpretation, only linear terms were considered, without interactions. The whole analysis was performed within the statistical software MATLAB®(MathWorks Inc., MA, USA).
For each agronomic trait, regression models at two crop stages were created using the field data from both years, following the image acquisition scheme presented in Table2. The first category of models was built for the crop stage with BBCH 25, which corresponds to two to three days before the first TdF (at the tillering stage). The second category was built for the crop stage with BBCH 45, which corresponds to approximately five days before the second TdF (at the booting stage). For the former category of models, the models were built considering the input variables from only the first image, captured at stage with BBCH 25. For the latter category of models, two approaches were tested. The first one used the union of input variables from two images, those captured at BBCH 25 and BBCH 45. The second also considers the input variables from the intermediate image acquisition at BBCH 31, which corresponds approximately 7–10 days after the first TdF (see the DAS values in Table2). The rationale for the latter was to include an intermediate input after the fertilization, in order to indirectly capture the effectiveness of the treatment, which could potentially increase the model’s predictive capabilities.
For each of the agronomic traits that changes during the season (plant height, N concentration, N uptake, and total biomass), one model was constructed for each stage considered, that is, for BBCH 25, 45, and 99 (harvested product, apart from plant height that does not change after booting). Obviously, for the models constructed before the second fertilization (BBCH 45, considering two or three images), predicting agronomic traits in previous stages (i.e., BBCH 25) has no practical use, since we actually try to predict a past state from the current status (which is not always possible). Hence, no models predicting agronomic traits at BBCH 25 were constructed at the booting stage. The models’ accuracy was assessed by means of the regression’s adjusted coefficient of determination (Adj.R2), as well as the root mean square error (RMSE). Since RMSE’s magnitude depends on the magnitude of the data, we also report a relative measure of the modeling errors’ dispersion, namely, the coefficient of variation of the RMSE (CVRMSE), expressed as a percentage:
CVRMSE=100·RMSE
y (%), (2)
whereyis the mean of the observed output values, that is, the mean of the field measurements for the agronomic trait the model is built for.
The Sequoia sensor incorporates a RE channel, which is an advantage for agronomic remote sensing studies, since it is sensitive to smaller changes of leaf health and plant stress in general.
However, many other sensors do not have such a channel, but rather follow the channel configuration of many high-resolution satellite sensors (blue, green, red, and NIR). Since it would be interesting to apply our methodology with such sensors as well, we also conducted a second experiment where the whole process (lasso input selection and linear model building) was repeated without the VIs relying on RE. In this case, the number of possible inputs was 23, namely, Sequoia’s green, red, and NIR channels and the following VIs: DVI, NDVI, SR, mSR, TNDVI, RDVI, SAVI, OSAVI, MSAVI2, gDVI, gNDVI, gRDVI, mSRG, GSAVI, MGSAVI, GWDRVI, CIG, MCARI1, MCARI2, and MTCARI (see Table3).
3. Results
During both rice cultivation periods, temperature and relative humidity were constantly recorded in order to compare the meteorological data sets. Table4reports monthly averages of temperature and relative humidity during the two growing seasons of the experimentation. The second year (2017) was slightly colder and less humid; however, these differences were minor and it can be concluded that they could not induce alterations in the morphophysiology of the rice plants between the two experiments over the two years.
Table 4.Monthly temperature and relative humidity during the growing seasons of the two years that the experiments were conducted. Minimum, maximum, and average values of each year are reported, averaged over each month.
Year Month Temperature (◦C) Relative Humidity (%) Minimum Maximum Average Minimum Maximum Average
2016
June 20 32 26 38 95 74
July 20 33 26 48 97 78
August 18 34 25 52 99 85
September 16 29 21 51 98 83
October 13 25 18 48 80 66
2017
June 18 32 25 50 84 61
July 19 34 26 53 96 73
August 18 34 24 60 98 79
September 14 32 21 52 97 78
October 13 28 19 43 89 69
Figure2 presents the differences of selected agronomic traits with respect to the fertilization treatment. Biomass, N uptake, and yield exhibit a rather expected trend, that is, higher values with increased amount of N fertilizer applied. Yield also exhibits the expected saturation with N fertilization, that is, the great differences in NU and BT between C2 (standard local practice) and C3 (double amount) at BBCH 45 were not fully reflected in yield differences. The variability was much higher at BBCH 25 (at tillering), since the canopy was not yet fully closed, due to the fact that the tillers were not completely developed and no TdF had been applied yet. As the growing season continued, the differences became more pronounced. The aforementioned trend was not clearly exhibited for N concentration and HI, which was also reflected in the modeling process that resulted in generally low predictive accuracy for those traits, as will be shown in the following. A possible explanation for this behavior is provided in the Discussion section. For completeness, Table5reports the results of a one-way ANOVA statistical analysis, followed by Tukey’s honest significant difference (HSD) post-hoc test [100]. The results of the latter are compactly presented by means of labeled groups (letters), with two treatments that include the same letter denoting statistically non-significant differences between the mean values of the corresponding agronomic trait (e.g., treatments C1 and C2 for the Yield trait were not found to exhibit statistically significant differences, whereas treatments C0 and C1 for the same trait did).
Table 5.Results of the one-way ANOVA test and Tukey’s honest significant difference (HSD) post-hoc test for the field-measured trait values against the different N treatments. The latter is presented by means of labeled groups, with two treatments that include the same letter denoting statistically non-significant differences between the mean values of the corresponding agronomic trait.
Trait p-Value for ANOVA F-test
Tukey’s HSD Group
C0 C1 C2 C3
PH25 0.3664 a a a a
PH45 1.62·10−20 c b b a
BT25 0.0918 a a a a
BT45 2.69·10−13 c b b a
BT99 3.59·10−13 c b b a
BSL99 8.25·10−6 b a a a
Yield 4.76·10−11 c b ab a
HI 0.8591 a a a a
NC25 0.0342 ab b ab a
NC45 0.0418 b ab ab a
NC99 0.1044 a a a a
NCSL99 0.0451 a a a a
NCG 0.0315 ab b ab a
NU25 0.0498 b ab ab a
NU45 1.31·10−18 c b b a
NU99 0.0002 b ab a a
NUSL99 0.0067 b ab a a
NUG 6.3·10−8 c b ab a
Remote Sens.2019,11, 545 11 of 25
Remote Sens. 2018, 10, x FOR PEER REVIEW 10 of 26
September 16 29 21 51 98 83
October 13 25 18 48 80 66
2017
June 18 32 25 50 84 61
July 19 34 26 53 96 73
August 18 34 24 60 98 79
September 14 32 21 52 97 78
October 13 28 19 43 89 69
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
(j) (k) (l)
Figure 2. Box plots of selected agronomic traits against the fertilization treatment: (a) BT25, (b) BT45, (c) Yield, (d) NU25, (e) N45, (f) NUG, (g) NC25, (h) NC45, (i) NUG, (j) PH25, (k) PH45, and (l) HI (see Table 1 for the description of agronomic traits).
Figure 2 presents the differences of selected agronomic traits with respect to the fertilization treatment. Biomass, N uptake, and yield exhibit a rather expected trend, that is, higher values with increased amount of N fertilizer applied. Yield also exhibits the expected saturation with N fertilization, that is, the great differences in NU and BT between C2 (standard local practice) and C3 (double amount) at BBCH 45 were not fully reflected in yield differences. The variability was much higher at BBCH 25 (at tillering), since the canopy was not yet fully closed, due to the fact that the tillers were not completely developed and no TdF had been applied yet. As the growing season continued, the differences became more pronounced. The aforementioned trend was not clearly exhibited for N concentration and HI, which was also reflected in the modeling process that resulted in generally low predictive accuracy for those traits, as will be shown in the following. A possible
Figure 2.Box plots of selected agronomic traits against the fertilization treatment: (a) BT25, (b) BT45, (c) Yield, (d) NU25, (e) N45, (f) NUG, (g) NC25, (h) NC45, (i) NUG, (j) PH25, (k) PH45, and (l) HI (see Table1for the description of agronomic traits).
Table6reports the accuracy measures (adjusted R2, RMSE, and CVRMSE) of all linear models constructed considering all 39 possible inputs (before input selection), along with the number of predictors (input variables) selected by means of the lasso approach. At the tillering stage (BBCH 25), high correlations were only observed for BT25 and NU25. Thus, this time point appeared to be too soon for predicting the future development of the plants, particularly, when the first TdF had not been applied yet. Nevertheless, BT25and NU25were considered the two most important traits to support N fertilization dose (units/ha) at tillering. The linear models constructed for the booting stage using two UAV images (at BBCH 25 and 45) exhibited high correlations for most of the agronomic traits.
The models for biomass (BT99and BSL99) and N concentration (NC99, NCSL99, and NUG) at maturity achieved relatively lower accuracies, but in most cases the adjusted R2values were higher than 0.7, with exception to HI. The linear models constructed for the booting stage, using three UAV images (at BBCH 25, 31, and 45), generally resulted in increased accuracies compared to the corresponding models constructed with only two images, although this was not always the case. The models that did exhibit an increase in accuracy always comprised a higher number of predictors than those with two images, which means that they exploited the additional information provided by the image at BBCH 31, as intuitively expected. Notable exceptions were the models for NUSL99and NUG, which resulted in higher accuracy with two images, but with overly complex models. In this case, the additional information provided by the image at BBCH 31 enabled the simplification of the models, but with a
penalty in accuracy at the same time. It is worth noting that for most of the agronomic traits (PH45, BT45, yield, NC45, and all NU traits), the models with only two images could achieve competitive performance. In all cases, the number of predictors selected by lasso was low, especially if we took into consideration the very high number of available predictors (see Section2.3). Specifically, the average number of predictors for all models was 4.39, with a median value of 4. For the models constructed for the tillering stage, 10.83% of the available predictors were used on average by the models, 5.77% for the models at booting with two images, and 3.85% for those at booting with three images. As such, the models could be easily visualized and the most important input variables could be identified, whereas the computational and—perhaps most importantly—storage requirements remained relatively low.
Table 6.Summary statistics of the derived linear models, built considering all 39 possible inputs (before input selection). Adjusted coefficient of determination (R2), root mean square error (RMSE), coefficient of variation of the RMSE (CVRMSE) (%), and number of predictors (#P) participating in each model are reported.
Trait Models at Tillering (1 image) Models at Booting (2 images) Models at Booting (3 images) Adj. R2 RMSE CVRMSE
(%) #P Adj.
R2 RMSE CVRMSE
(%) #P Adj. R2 RMSE CVRMSE
(%) #P
PH25 0.12* 2.59 8.95 1 — — — — — — — —
PH45 0.50 5.97 9.74 2 0.84 3.43 5.59 3 0.84 3.41 5.57 2
BT25 0.86 0.43 23.09 7 — — — — — — — —
BT45 0.56 1.03 22.51 3 0.87 0.55 12.01 2 0.87 0.55 12.01 2
BT99 0.39 3.17 23.12 1 0.74 2.07 15.07 3 0.74 2.07 15.07 3
BSL99 0.54 1.99 25.89 3 0.72 1.54 20.07 3 0.72 1.55 20.23 4
Yield 0.61 1.47 19.12 4 0.77 1.13 14.73 5 0.80 1.06 13.73 6
HI 0.24 0.05 9.21 1 0.31 0.04 8.75 1 0.31 0.04 8.75 1
NC25 0.16 0.26 7.44 1 — — — — — — — —
NC45 0.68 0.32 14.23 3 0.88 0.20 8.87 6 0.88 0.20 8.67 7
NC99 0.77 0.22 17.67 5 0.70 0.25 19.98 3 0.85 0.18 14.29 6
NCSL99 0.70 0.28 30.97 5 0.86 0.20 21.49 11 0.80 0.23 25.43 6
NCG 0.57 0.08 6.49 3 0.64 0.08 5.95 3 0.72 0.07 5.24 4
NU25 0.80 19.70 29.84 3 — — — — — — — —
NU45 0.50 31.88 30.51 3 0.86 16.72 16.00 3 0.86 16.72 16.00 3
NU99 0.68 60.63 33.10 4 0.69 59.24 32.34 3 0.75 53.94 29.45 4
NUSL99 0.68 40.20 48.97 5 0.93 18.48 22.51 12 0.82 29.85 36.37 7
NUG 0.58 24.79 24.53 5 0.75 19.02 18.81 5 0.83 15.89 15.72 8
* Statistically significant at a level of 0.05; all other models are statistically significant at a level of 0.01.
In order to give a visual representation of the models’ errors, Figure3depicts the scatter plots of predicted versus observed values for some indicative models constructed in each stage. The gray dashed line represents the ideal perfect linear relationship, whereas the continuous red line is the simple linear regression line of the predicted versus observed data. Models with high adjusted R2 values exhibited, generally, uniform distribution of data around the ideal prediction line, with the regression line being very close to the latter and very few outliers observed in a few cases. For completeness, the scatter plots for all models constructed are provided as Supplementary Materials.
Table7presents the linear models constructed at the tillering stage, whereas Tables8and9present the models constructed at the booting stage using features extracted from two and three UAV images, respectively. For convenience, the adjusted R2and RMSE values of Table6are replicated in these tables as well. Focusing on the models constructed at the tillering stage that exhibit high accuracy (BT25, NC99, and NU25in Table7), they mostly exploited chlorophyll-sensitive vegetation indices (gRDVI, CIG, CIRE, TCARI, MCARI, MRETVI) together with the green and red reflectance channels. At the BBCH 25 stage, the canopy was not fully closed, introducing strong background soil/water effects.
This in turn decreased the efficiency of traditional biomass-sensitive VIs (e.g., NDVI, SR, and mSR) for estimating agronomic traits at the tillering stage. Conversely, chlorophyll-sensitive VIs incorporated the green or RE channel (or both), which rendered them more sensitive to small variations of biomass.
The latter was also true for the model built for yield, although the latter exhibited relatively lower accuracy (Adj. R2= 0.61; RMSE = 1.47 tn·ha−1) than the aforementioned ones.
